Your handicap is not your average

A Handicap Index is the average of the best eight of your last twenty rounds. Everyone repeats that sentence and almost nobody finishes the thought: a best-of-eight average is not an estimate of how you play. It is an estimate of how you play on a good day, and the gap between the two is not small, not constant, and not the same for you as it is for the person you are playing.

What the number actually computes

Under the World Handicap System, every round you post becomes a Score Differential — your adjusted gross score, corrected for the course’s rating and slope, so a 92 at a brutal course and an 84 at an easy one can sit in the same column. Your Handicap Index is the mean of the lowest eight of the most recent twenty differentials. Twelve of your twenty rounds are thrown away, and they are always the same twelve: the worse ones.

That rule has a fixed statistical consequence. If your differentials scatter with standard deviation σ, the average of the lowest eight of twenty sits, on average, 0.93 σ below your mean. Nothing about your swing is involved. So your index is not your average; it is roughly your average minus one standard deviation, and the only way to have a small gap is to be very consistent.

So I ran it on 229,333 real rounds

Tour golf is the only place I can get a long, clean, public record of round-by-round scores, so that is where I tested it. Everything in this section is about professionals; the amateur numbers come later and are worse. I pulled every first and second round from the PGA Tour, LPGA, DP World Tour, Korn Ferry Tour and PGA Tour Champions from 2021 to September 2026 — 229,333 rounds by 4,894 players across 916 events — and turned each one into a differential the way the WHS does, using the field’s scoring average that day in place of a course rating. Then, for every player with a long enough record, I built a Handicap Index from a rolling window of twenty rounds and compared it with the round they actually played next. That is 180,204 predictions over 1,482 players.

The index came in 2.60 strokes below what those players averaged. They beat it in 18.4% of rounds. Simulating the same rule on normal scores — including the fact that the index is itself estimated from twenty rounds, which the textbook 17.7% figure leaves out — predicts 18.5%. Dean Knuth’s odds table, published by the USGA long before the WHS existed, puts a player at or under their handicap at 5:1 to 6:1 in tournament play, which is one round in six to one in seven. Three sources, all landing between roughly 14% and 19%, none of them anywhere near the half you would get from an average.

The gap is not a tour artefact and it is not occasionally negative. Across all 1,482 players the median gap is 2.57 strokes, the smallest is 1.54, the largest is 4.04, and not one player has a gap of zero or less. They cannot. A best-of-eight average is bounded above by an average of twenty.

The gap is not a constant, it is your spread

If the gap were the same for everyone, none of this would matter — you could subtract 2.6 and carry on. It is not. Regressing each player’s gap on how much their rounds scatter gives

gap = 0.84 × (standard deviation of their rounds) + 0.23     r = 0.95

which is the 0.93 from the theory, attenuated because the x-axis is itself a noisy twenty-round estimate of the spread. Ninety percent of the variation in how much your index flatters you is explained by how streaky you are.

That means the index systematically confuses two different things. Paul Casey and Kevin Na carry an identical index in this data — −3.536 and −3.537 — and they are not the same golfer. Across his twenty-round windows Casey scatters with a standard deviation of 2.00 and averages 1.70 strokes better than the field. Na scatters 3.40 and averages 0.43 better. Same index, 1.27 strokes apart. I went looking for that pair — it is the closest-index, widest-spread-difference match on the PGA Tour in this data — but the effect it illustrates is in every row.

Scale it up and the ranking breaks. Take any two players whose indexes sit within half a stroke of each other, and 27% of the time the lower index belongs to the player with the worse average (25.6% if I keep only players with at least thirty windows). Tighten the window to a tenth of a stroke and it is 45%, which is nearly a coin flip. Some of that is noise in the yardstick — a player’s average is estimated too — but the direction is not in doubt, and the mechanism is the one in the chart above: the index is the average minus a spread it never reports.

Amateurs have it worse

Knuth’s odds table is published per handicap bracket, and you can invert it: find the normal distribution whose tail probabilities match each row, and you recover the spread for that bracket. That gives a standard deviation rising from 2.60 strokes for the 0–5 bracket to 5.29 for 30-plus. I take only the spread from that table and nothing else, because the table predates the WHS — its rows are measured against the old best-10-of-20 × 0.96 index, and against tournament rounds, so its implied gaps belong to a system that no longer exists.

Running the current best-of-eight rule at those spreads gives the gap directly: 1.9 strokes at scratch, 3.1 at a 15, 4.7 at a 36. The beat rate does not move — about 18.5% at every level — because best-of-eight on normal scores is scale-free, so the rule cannot fail to produce that. What scales is the number of strokes, and strokes are what you lose the bet by.

Independently, MyGolfSpy’s study of TheGrint’s database found low handicaps averaging four to five strokes above their index and high handicaps six to eight — larger than my figures because gross score versus index also carries the course rating, and because casual rounds scatter more than tournament ones. Different data, same direction, same slope.

The academic work landed here thirty years ago and nobody moved. Bingham and Swartz, in Equitable Handicapping in Golf, fit differentials as normal, estimated σ(I) = 2.74 + 0.053I from 49 golfers’ records, and showed that in matches decided by good rounds the better golfer won only 34% to 36% of the time. Swartz’s follow-up proposal for the Royal Canadian Golf Association used the best sixteen of twenty with order-statistic weights and got 46% of golfers shooting their handicap instead of 11%.

It is also a worse forecast than the obvious alternative

Bias you can argue about. Prediction you cannot. I gave seven rules the same twenty rounds and asked each to guess the twenty-first, 180,204 times:

The plain mean of twenty is 27% more accurate than the index, and it is worth being precise about where that 27% comes from: almost all of it is the bias. Recentre the index by adding 2.60 back and it gets to 2.31, within about 1% of the mean. What is left is real but small — eight rounds is a noisier estimate of a centre than twenty are — and the honest summary is that the index’s problem is overwhelmingly where it points, not how tightly.

The other thing in that table is a result I expected to go the other way. Recency weighting buys nothing: an eight-round half-life and a flat mean of twenty finish within 0.005 strokes of each other, which on the overlapping-window caveat above is no difference at all. Form, at the granularity of twenty rounds, is mostly a story we tell.

Why the rule exists anyway

The best-of-eight rule is not stupid, it is optimised for something other than accuracy, and it is worth saying what before replacing it.

It is a potential measure on purpose. A handicap is meant to let anyone compete with anyone on a given day, and on a given day the thing that decides the match is whether someone finds their good round. Setting the bar near a player’s good round makes beating it feel like an achievement and makes a net 68 mean roughly the same thing at every level.

It resists sandbagging in one direction. If your index were your average, posting a few 100s would move it a lot. Under best-of-eight, bad rounds are invisible — you cannot inflate your index by playing badly, only by playing badly on the eight rounds that happen to be your best, which is hard. The WHS adds a soft cap at 3.0 strokes above your Low Handicap Index, a hard cap at 5.0, and an automatic reduction for scores 7.0 better than your index, all to police the other direction.

And it is one number, printed on a card, usable by a pro shop clerk. That constraint is doing more work than any statistical argument.

Four numbers that would do the job better

1. Expected Score: the mean of all twenty. The single highest-value change, and the cheapest — every system already holds the twenty differentials. It is unbiased by construction, it stops punishing consistency, and it makes “I shot my handicap” mean what people think it means. Keep the net-double-bogey cap on each hole so a single blow-up hole cannot swing it, and the mean is as robust as the best-of-eight was. If you distrust blow-up rounds specifically, the trimmed mean that drops the four best and four worst finished within 0.01 strokes of the plain mean here, and would hold up better on amateur records, which are far more right-skewed than tour records.

2. A second number: the spread. Print 14.2 ± 3.6 instead of 14.2. It costs one extra column in the database and it is the number that resolves every ambiguity above: it tells you Casey and Na are not the same golfer, it tells a committee which players to look at, and it tells you how much to trust the first number. The WHS is already thinking about spread — Appendix D’s handicap review tool compares a player against an expected standard deviation and an expected scoring range — but that expectation is derived from the index itself and is therefore identical for every player carrying it. The player’s own observed spread is sitting right there in the same twenty rows, and nobody computes it.

3. Shrinkage when the record is short. The WHS lets you post an index off three scores, using the lowest one minus 2.0. That is not just noisy, it is biased: my simulation of a true 15-handicap gets a mean three-score index of 13.3, 1.7 strokes low, with a standard deviation of 2.5 and a 90% range from 9.0 to 17.2 — against 13.5 to 16.5 once twenty scores are in. Ask each rule to predict the next round: the three-score WHS index is off by 5.3 strokes, the plain mean of the three by 3.0. Almost all of that win is the mean, not anything clever. Shrinking that mean toward the population adds only another 0.04 strokes here, because golfers differ from each other far more than three rounds of noise does — but it costs four lines of arithmetic, it stops a brand-new index from being wild, and it is what every other rating system does with sparse data.

4. For competition, normalise by the spread. This is the one that cannot be fixed with a better centre, and it is the strongest argument in the literature. Take two simulated 15-handicaps, one scattering 2.5 and one scattering 5.0. In a single net round off their equal indexes, the steady player wins 66% of the time, because they are genuinely 2.3 strokes better. In a five-round competition where only the best round counts, the same steady player wins 45% — the streaky player’s tail wins tournaments. Switch both to playing off their own average and the single round becomes fair at 49.9%, but the tournament gets worse, at 22%. Only Swartz’s fix works in both: score each round as (differential − your average) ÷ your spread, a z-score rather than a stroke count, and the win rate is 50.0% either way.

What to do with this

• If you track your own rounds, compute the mean of your last twenty differentials alongside your index. That number, not the index, is what to expect on Saturday. On these figures it will sit roughly 2 strokes above your index if you are a scratch player and closer to 5 if you are a 36.
• Compute your standard deviation too, and watch it. Getting more consistent shows up there first; it shows up in your index only slowly, and it shows up in your average as a rise in the gap, which looks like nothing.
• Stop reading a beaten handicap as a breakthrough. Beating it in about one round in five and a half is the system working exactly as designed. Beating it eight times in a twenty-round record happens about 2% of the time by chance — that one is worth taking seriously.
• When you set up a side game, agree which kind of contest it is. For a single round, strokes off each player’s average is the fairer allowance. For anything decided by a best round out of several, use full indexes and accept that you have chosen a format that rewards streakiness.
• If you are building anything on top of handicaps — a matchmaking system, a betting line, a pace-of-play model — do not feed it the index. Feed it the mean and the spread. The index is an order statistic wearing a point estimate’s clothes, and it will bias everything downstream in a way that scales with how inconsistent your users are.